Fluid network abnormal source positioning and knowledge integration method and system

Abstract

The invention belongs to the technical field of fluid network anomaly positioning, and particularly relates to a fluid network anomaly source positioning and knowledge integration method and system. Comprising the following steps: S1, collecting basic data of a fluid network, including a topological relation, branch parameters and fluid data; s2, randomly selecting a fault branch on the fluid network, setting a fault type and a fault degree, randomly generating fault wind resistance, and performing simulation calculation on pressure data, flow data and concentration data before and after abnormity according to the wind resistance before and after the fault; s3, fluid network calculation is divided into a plurality of sub-networks; s4, calibrating normal and abnormal labels of the pressure data, the flow data and the concentration data; s5, training the training data set to obtain an optimal abnormal source positioning model; s6, collecting fluid data of the fluid network through a sensor, and outputting whether abnormity occurs or not and an abnormal position; and S7, continuously repeating the steps S1 to S6. According to the invention, accurate and efficient positioning of the abnormal source is realized through subnet division and multi-module cooperation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fluid network anomaly positioning, and particularly relates to a fluid network anomaly source positioning and knowledge integration method and system. BACKGROUND

[0002] Fluid networks exist widely in many fields such as mining, chemical industry, petroleum, natural gas, water conservancy, medical treatment, environmental protection and urban lifeline system. A fluid network usually involves complex pipeline systems, valves, pumps, sensors and other equipment and facilities, which together constitute a complex fluid network system and are responsible for on-demand and quantitative delivery of various fluid media. However, due to various factors such as equipment aging, improper operation and external interference, various abnormal phenomena such as leakage, blockage and pressure fluctuation may occur in the fluid network during operation. These abnormal phenomena not only affect the normal operation of the fluid network, but also may have a serious impact on industrial production safety and the like. Therefore, it is particularly important to accurately locate and timely handle the abnormal source of the fluid network. The existing methods generally cannot realize the generalization and integration of the abnormal source positioning knowledge of a complex network, and cannot realize the migration of the abnormal source positioning model of the fluid network of different network scales and topological structures through self-learning. SUMMARY

[0003] The application provides a fluid network anomaly source positioning and knowledge integration method and system to solve the problem that there is no generalizable abnormal source positioning model in the field of fluid networks and the self-learning and migration of the fluid network suitable for different scales and structures cannot be realized.

[0004] The application adopts the following technical scheme: a fluid network anomaly source positioning and knowledge integration method, comprising the following steps: S1: collecting fluid network basic data, including topological relationship, branch parameter and fluid data; topological relationship: branch adjacency matrix A E , node and branch association matrix B; branch parameter: branch length l, diameter d, friction coefficient f and cross-sectional area s; fluid data: pressure, flow rate and concentration data measured by a sensor; S2: randomly selecting a fault branch on the fluid network and setting a fault type and a fault degree, and randomly generating a fault wind resistance amount , simulating and calculating the pressure data, flow rate data and concentration data before and after the anomaly according to the wind resistance before and after the anomaly; S3: dividing the calculation of the fluid network into a plurality of sub-networks for knowledge integration and abnormal positioning reasoning of the model; S4: labeling the normal and abnormal labels of the pressure data, flow rate data and concentration data, and storing the fluid data, sub-network data and label data into a database as a training data set. S5: training the data set to obtain the optimal abnormal source positioning model for fluid network abnormal source positioning; S6: collecting fluid data of the fluid network through the sensor, outputting whether an abnormality occurs and the abnormal position, verifying and processing the abnormality; S7: continuously repeating S1-S6 to realize real-time positioning of the abnormal source of the fluid network and dynamic generalization integration of the abnormal source positioning knowledge.

[0005] In step S2, the mathematical model of the simulation calculation is: In the formula, Q is the simulated flow of all the tunnels, composed of q1, q2,..., qn, m / s; P is the simulated pressure of all the nodes, composed of p1, p2,..., pna; c is the simulated gas concentration of all the tunnels, composed of c1, c2,..., cn, %; A is the tunnel adjacency matrix; P' is the perceived pressure matrix; Q' is the perceived flow matrix; c' is the fluid concentration matrix, used in the mixed fluid flow network; d is the tunnel pipe diameter matrix, m; f is the tunnel friction coefficient matrix; l is the pipe length matrix, m; and F is the network solving algorithm. n 3 m E The calculation model of the network solving algorithm F is as follows: wherein, L 1 represents the node mass conservation equation; L 2 represents the node component conservation equation, used in the mixed gas flow network; L 3 represents the resistance law; r ij represents the branch air resistance between node i and node j; L 4 represents the density expression of node i based on the gas state equation; and L 4 is substituted into L1 and L 2 , to obtain the fluid network model L = (L 1 ,L 2 ,L 3 ) The simulation calculation result is obtained by iteratively solving the fluid network model L by the Newton iteration method.​​​​

[0006] Step S3 comprises: S31: For fluid network , is a set of nodes, ; is a set of lanes, ; AE is a lane adjacency matrix, when the lane is adjacent to the lane , , otherwise ; The random lane access probability is obtained by the following formula: In the formula, denotes the number of adjacent lanes of the lane ; denotes the sum of the number of adjacent lanes of all lanes; denotes the probability of randomly accessing the lane ; S32: Randomly select a core lane based on the random walk algorithm , and include the core lane together with its two end nodes , into the sub-network; at the same time, the first-order neighborhood , and the adjacent edges of the two end nodes are also included in the sub-network; finally, a star topology structure TS(ex) is formed with the core lane as the center, and the core lane and its associated local network together constitute the structural main body of the sub-network.

[0007] Step S4 comprises: S41: According to the simulation calculation results under normal and abnormal conditions in step S2, the fault label calibration is performed on the pressure, flow rate and concentration data; S42: The fluid data under normal and abnormal conditions obtained in step S2, the sub-network obtained in step S3, and the corresponding labels are stored into the database as training data.

[0008] Step S5 comprises: S51: Topological feature integration, based on the fluid network topological relationship, a first-order neighborhood prior feature integration with weights is constructed; S52: Data feature integration, based on the fluid network data, a first-order neighborhood posterior feature integration with weights is constructed; S53: Autonomous learning and training of the abnormal positioning model.

[0009] Step S52 comprises: The cross-entropy loss function for fault diagnosis is constructed: In the formula: is the cross-entropy loss function of fault diagnosis; C g is the number of categories; is the true label of the fault; is the predicted label of the fault; is the topological feature, i.e., the lane adjacency matrix of the ventilation network; X is the data feature; The data feature, topological feature, and fault label of the subnetwork are used as the input of the model; After topological feature integration and data feature integration, the features are connected, and the Softmax layer is used to calculate the predicted label. The label dimension is the number of lanes x the number of fault types. The cross-entropy loss of the predicted label and the true label is calculated, and the loss is backpropagated to realize autonomous learning, thereby realizing autonomous learning training of the abnormal positioning model.

[0010] A fluid network abnormal source positioning and knowledge integration system, comprising: A subnetwork calculation module for dividing the fluid network into several subnetworks in step S3; A simulation module for simulating in step S2; A knowledge integration module for autonomous learning training of the abnormal source positioning model in step S5; An abnormal positioning module for positioning the abnormal source of the fluid network in step S6; A data management module for importing, storing, exporting, generating, and calibrating the topological relationship in step S1, the fluid data in step S2, and the label data in step S4.

[0011] The subnetwork calculation module includes a random subnetwork dynamic calculation module and a fixed subnetwork module; The random subnetwork dynamic calculation module obtains several subnetworks; The fixed subnetwork module obtains fixed subnetworks through manual input.

[0012] The knowledge integration module includes: A first data import module that receives the stored topological relationship, fluid data, and label data in the data management module; A feature integration module for completing topological feature integration and data feature integration; An autonomous learning module for training the abnormal source positioning model.

[0013] The abnormal positioning module includes: a second data import module, which imports fluid data, topological relations and sub-networks from the data management module as inputs of the abnormal positioning model; a knowledge model module, which is the optimal abnormal source positioning model trained by the knowledge integration module; an abnormal output module, which is responsible for finally transcoding and outputting the position of the abnormality according to the calculation result of the optimal abnormal source positioning model.

[0014] Compared with the prior art, the present application has the following beneficial effects: The fluid network abnormal source positioning and knowledge integration method provided by the present application realizes accurate and efficient positioning of the abnormal source through sub-network division and multi-module cooperation, and significantly improves the positioning speed and emergency response capability in combination with real-time data acquisition and simulation technology. The knowledge integration module based on the autonomous learning and back propagation mechanism can optimize the model parameters by minimizing the loss function, and supports model migration across network scales and topological structures, greatly enhancing the generalization ability. The data management module built in the system comprehensively integrates topological, fluid and label data, supplemented by dynamic simulation data, ensuring data integrity and training effect. The method can be widely applied to tree-shaped or network-shaped fluid networks in the fields of mining, chemical industry, energy and medical treatment, and is compatible with multiple scenes such as gas extraction, mine ventilation and medical gas supply, greatly reduces the labor cost through automatic positioning, improves the safety of the system, and provides efficient and reliable technical support for industrial production and urban lifelines. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 It is a fluid network abnormal source positioning and knowledge integration method architecture diagram; Figure 2 It is a feature integration schematic diagram; Figure 3 It is a gas extraction system schematic diagram of the embodiment of the present application; Figure 4 It is a mine ventilation system schematic diagram of the embodiment of the present application; Figure 5 It is a medical gas transportation schematic diagram of the embodiment of the present application; In the figure: 3-1 - extraction pump station; 3-2 - extraction pipeline; 3-3 - extraction node; 3-4 - drilling field; 3-5 - coal seam; 3-6 - control valve; 4-1 - ventilator; 4-2 - roadway; 4-3 - air door; 5-1 - pressure pump; 5-2 - transportation pipeline; 5-3 - transportation terminal; 5-4 - valve. DETAILED DESCRIPTION

[0016] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0017] As shown in Figure 1 , a fluid network anomaly source positioning and knowledge integration system comprises: a sub-network calculation module for dividing a large fluid network into a plurality of sub-networks; a simulation and emulation module for simulation and emulation of fluid network data; a knowledge integration module for autonomous learning and training of an anomaly source positioning model; an anomaly positioning module for positioning of a fluid network anomaly source; a data management module for import, storage, export, generation and calibration of fluid network topology data, fluid data and label data.

[0018] As shown in Figure 1 , it is an architecture diagram of the fluid network anomaly source positioning and knowledge integration system.

[0019] The sub-network calculation module comprises a random sub-network dynamic calculation module and a fixed sub-network module, and the sub-network is calculated through input fluid network topology data. The random sub-network calculation function obtains a plurality of sub-networks through a random walk algorithm on the fluid network; the fixed sub-network module inputs the fixed sub-network through manual input.

[0020] The random walk algorithm generates a star-shaped topology network structure T S by accessing different branches, taking the accessed branch as the core. For a fluid network , is a node set, ; is a branch set, ; AE is a branch adjacency matrix, when the branch is adjacent to the branch , , otherwise . The random branch access probability is obtained by the following formula: In the formula, the number of adjacent branches of the roadway is represented by ; the sum of the adjacent branch numbers of all roadways is represented by ; and the probability of randomly accessing the roadway is represented by .

[0021] Core tunnels are selected based on random probability. Connect the core tunnel with its two end nodes , Incorporate them into the subnetwork; at the same time, include the first-order neighborhoods of the nodes at both ends. , The adjacent lanes are also included in the sub-network; ultimately forming a network centered on the core lanes. The core is a star-shaped topology TS(ex), and the core lanes and their associated local networks together constitute the main structure of the subnetwork.

[0022] The simulation module is used to simulate fluid network data. When the amount of data input to the knowledge integration module is insufficient, it can be used to supplement the training dataset. The mathematical model of the simulation module is as follows: In the formula, To simulate the flow rates of all roadways, we have q1, q2, ..., q n Composition, m 3 / s; P represents the pressure at all nodes obtained from the simulation, defined by p1, p2, ..., p m Composition, pa; c represents all simulated roadway gas concentrations, used in mixed fluid flow networks, composed of c1, c2, ..., cn, %; A E denoted as _i ...

[0023] The computational model of the network solution algorithm F is as follows: in, L 1 Represent the equation for the conservation of nodal mass; L 2 This represents the nodal component conservation equation, used in mixed gas flow networks; L 3 This represents the law of resistance; r ij This represents the branch resistance between node i and node j; L 4 This represents the density expression for node i obtained based on the gas law. L 4 Substitute L1 and L 2 In this process, a fluid network model is obtained. L = (L1 ,L 2 ,L 3 ) The simulation calculation result is obtained by iteratively solving the fluid network model through Newton iteration method. L

[0024] The knowledge integration module includes data import, feature integration and autonomous learning functions. The data import is responsible for importing the topological relationship data, fluid data and label data of the fluid network or fluid sub-network from the data management module; the feature integration module includes topological feature integration and data feature integration, and the topological features and data features of the fluid network are considered, so that the abnormal positioning model can be generalized and migrated, as shown in FIG. 2. Figure 2 The autonomous learning completes the back propagation of the positioning model parameters by minimizing the loss function, and realizes the autonomous learning training of the abnormal positioning model.

[0025] The topological feature integration is a first-order neighborhood prior feature integration based on the weight of the topological relationship of the fluid network, and is calculated as follows: In the formula, 1 represents a topological feature integration matrix; A 1 represents a topological feature integration matrix; B 1 represents a topological feature integration matrix; D 1 represents a topological feature integration matrix; I 1 represents a topological feature integration matrix; A E 1 represents a topological feature integration matrix. F 0 represents a fluid network feature; W 0 represents a fluid network feature; δ 0 represents a fluid network feature; F 1 represents a topological feature integration matrix;

[0026] The data feature integration is a first-order neighborhood posterior feature integration based on the weight of the data of the fluid network, and is calculated as follows: In the formula, 2 represents a data feature integration matrix; A 2 represents a data feature integration matrix; W 2 represents a data feature integration matrix; W 2 represents a data feature integration matrix; θ 2 represents a data feature integration matrix; f 2 represents a data feature integration matrix; F 2 represents a data feature integration matrix; i 2 represents a data feature integration matrix; j 2 represents a data feature integration matrix;​k Representative label.

[0027] The anomaly positioning module uses the knowledge integration module to finally generate an anomaly source positioning knowledge model M for anomaly source positioning, including data import, knowledge model, anomaly output, and anomaly verification functions. Among them, the data import imports real fluid data, topology data, and subnet data from the data management module as the input of the anomaly positioning model; the knowledge model is the optimal anomaly source positioning model M trained by the knowledge integration module, and the essence of the model M is the encapsulation of the feature integration process; the anomaly output is responsible for finally transcoding and outputting the position of the anomaly according to the calculation result of the model M; the anomaly verification is to verify whether the anomaly position output by the anomaly source positioning model is truly abnormal by the on-site staff or other means, and to use the data management module to manually label the real anomaly position.

[0028] The mathematical model of the anomaly positioning module is represented as follows: In the formula, represent the fault prediction label output by the model; M represents the optimal anomaly source positioning model trained; AE represent the roadway adjacency matrix; 、 represent the real pressure data perceived at t+1 and t time respectively; represent the real flow data perceived at t+1 and t time respectively; represent the real concentration data perceived at t+1 and t time respectively, which is used in the mixed fluid flow network; TS is the subnet data; d is the roadway pipe diameter matrix; f is the friction coefficient matrix; l is the pipe length matrix; s is the pipe section area matrix.

[0029] The data management module includes data acquisition, label labeling function and data storage function. Among them, data acquisition includes real fluid data integration (through sensors) and simulation fluid data generation (through simulation calculation); label labeling is to label the corresponding network or subnetwork branch of a group of fluid data as normal label or abnormal label respectively, and the labeled data is archived in the data storage for training of the anomaly source positioning model. The data storage function is to save the data set including topology data, fluid data, label data and subnet data, which is used for data import of the knowledge integration module and the anomaly source positioning module.

[0030] Example 1: Gas extraction network anomaly positioning As Figure 3As shown, a gas drainage network typically has a tree-like structure, consisting of drainage pump stations 3-1, drainage pipelines 3-2, control valves 3-6, monitoring and control systems, and other auxiliary devices. Drilling sites 3-4 are located within coal seams 3-5, and different drilling sites 3-4 connect to drainage nodes 3-3. The fluid in the pipeline network is a binary mixture of gas and air. During operation, the gas drainage network may experience blockages, leaks, and other abnormalities. These abnormalities require accurate and timely location and handling to ensure the safety of underground workers. This embodiment will achieve knowledge integration and location of abnormal conditions through the following process. Figure 3 This is a simplified diagram of a gas extraction system.

[0031] S1: Collect basic data of the gas extraction network, including topology and fluid data.

[0032] In this embodiment, the acquired data includes: pipe adjacency matrix A E The following are the relationships between branches and nodes: B, real fluid data (including sensor-sensed pressure matrix P', flow matrix Q', and gas concentration matrix c'), pipe length matrix l, pipe diameter matrix d, pipe friction coefficient matrix f (including valve friction coefficient), and pipe cross-sectional area matrix s.

[0033] See the simplified diagram of the gas extraction system. Figure 3 Gas extraction network diagram , For a set of nodes, ; For the collection of alleyways, ; AE is the lane adjacency matrix, when the lane and alleyway When adjacent, ,otherwise .

[0034] S2: Randomly select the fault branch and set the fault type and fault degree, and randomly generate the fault wind resistance. Based on the wind resistance before and after the fault, perform simulation calculations on the initial pressure, fault pressure, initial flow rate, fault flow rate, initial concentration, and fault concentration before and after the anomaly.

[0035] In this embodiment, simulation calculations are performed on the fluid data prior to the anomaly: In the formula, To simulate all pipe flow rates, given q1, q2, ..., q 12 Composition, m 3 / s; P represents the pressure at all nodes obtained from the simulation, defined by p1, p2, ..., p 13 Composition, pa; c represents the simulated gas concentrations in all roadways, composed of c1, c2, ..., c12 Composition, A E is the pipe adjacency matrix; P' is the perceived pressure matrix; Q' is the perceived flow matrix; c' is the perceived gas concentration matrix; d is the roadway pipe diameter matrix, m; f is the roadway friction factor matrix; l is the pipe length matrix, m; and F is the network solution algorithm.

[0036] In this embodiment, the abnormal fluid data is simulated and calculated, First, different types and degrees of abnormalities are set for the random branches of the gas extraction network: In the formula, C represents the gas network abnormality setting matrix; E represents each branch, ; the X vector represents the abnormal type, In the gas extraction network, xi=0 represents no abnormality, xi=1 represents leakage, and xi=2 represents blockage; the A vector represents the abnormal degree, , The value range is 0~1; Then, the fluid data under abnormal conditions is simulated and calculated: The gas fluid network solution calculation model is as follows: wherein, L 1 represents the node mass conservation equation; L 2 represents the node gas component conservation equation; L 3 represents the resistance law; r ij represents the branch air resistance between node i and node j; L 4 represents the density expression of node i based on the gas state equation. Substitute L 4 into L 1 and L 2 , to obtain the gas fluid network model L = (L 1 ,L 2 ,L 3 ) The simulation calculation result is obtained by iteratively solving the fluid network model L by the Newton iteration method.

[0037] S3: Use the sub-network calculation module to calculate the sub-networks of different fluid networks, for knowledge integration and abnormal positioning reasoning of the model.

[0038] Based on the random walk algorithm, randomly access the branch of the pipe segment, and take the accessed random branch as the core branch e x , together with its two end nodes x , , , and the first-order neighborhood of the two end nodes , and the adjacent edges, form a star topology T x centered on the core branch e S (e x ).

[0039] In this embodiment, the calculation formula of the random access probability of the pipe branch is as follows: In the formula, represents the number of adjacent roadways of the pipe ; represents the sum of the numbers of adjacent pipes of all pipes; and represents the probability of randomly accessing the pipe.

[0040] The random pipe access probability of this embodiment is as follows: .

[0041] S4: Use the data management module to label the normal and abnormal labels of the data, and store the fluid data, sub-network data in the normal and abnormal states and their corresponding abnormal labels into the database as training data.

[0042] According to the simulation calculation results in the normal and abnormal states in S2, the pressure, flow rate, and concentration data are labeled, and the fluid data in the normal and abnormal states obtained in S2, the sub-networks obtained in S3, and the fault labels are stored into the database as training data.

[0043] S5: The knowledge integration module reads the training data set, and trains to obtain the optimal abnormal source positioning model M through feature integration and autonomous learning. The optimal abnormal source positioning model M is updated to the knowledge model of the abnormal positioning module, and is used for fluid network abnormal source positioning through centralized deployment.

[0044] In this embodiment, the topological feature integration and data feature integration are respectively performed by the following formulas.

[0045] In the formula, A1 represents a topological feature integration matrix; B represents a branch and node association matrix; D represents a node degree matrix; I represents an identity matrix; A E represents a branch adjacency matrix. F 0 represents a fluid network feature; W 0 represents a learnable parameter matrix; diag represents diagonalization; δ represents a nonlinear transformation; F 1 represents integrated features.

[0046] In the formula: A 2 represents a data feature integration matrix; W , W 1, θ represents a learnable parameter matrix; f represents a certain column of a feature matrix; F 2 represents integrated features; i , j , k represents a label.

[0047] S6: Collect gas extraction network fluid data through sensors, import into the abnormal positioning module in real time through the data management module, output whether an abnormality occurs and the abnormal position, verify and handle the abnormality. Upload and store the data and labels, and update the database.

[0048] In this embodiment, the abnormal label prediction is performed by the following formula: In the formula, represents a fault prediction label output by the model; M represents an optimal abnormal source positioning model obtained by training; AE represents a pipeline adjacency matrix; respectively represent the real pressure data perceived at t+1 and t; respectively represent the real flow data perceived at t+1 and t; respectively represent the gas concentration data perceived at t+1 and t; TS is a subnetwork data; d is a roadway pipe diameter matrix; f is a friction coefficient matrix; l is a pipeline length matrix; s is a pipe section area matrix.

[0049] Then, the obtained abnormal label prediction is verified, and the verified data and labels are transmitted to the data management module to update the training database.

[0050] S7: Repeat S1~S6 continuously to achieve real-time location of abnormal sources in the gas extraction network and dynamic generalization integration of abnormal source location knowledge.

[0051] Example 2: Ventilation Network Anomaly Location Mine ventilation networks are typically complex mesh structures used for the circulation and exhaust of mine air, such as... Figure 4 As shown, the system includes a ventilation fan 4-1, a roadway 4-2, and an air door 4-3. Aging of roadway 4-2, roadway collapse, and abnormal opening and closing of the air vents in air door 4-3 can cause changes in roadway air resistance, leading to changes in airflow in localized areas or even the entire mine. This disrupts the stability and reliability of the ventilation system; such faults are called resistance-change faults. Resistance-change faults directly affect the correct operation of the optimization decision-making algorithm, leading to increased algorithmic bias. Therefore, resistance-change fault diagnosis is a key aspect of ventilation network anomaly detection. To effectively locate abnormal opening / closing of air doors and roadway collapse anomalies, this embodiment will integrate and locate anomaly knowledge through the following process.

[0052] S1: Collect basic data on the mine ventilation network, including topology and fluid data.

[0053] In this embodiment, the acquired data includes: the lane adjacency matrix A E The system includes: the correlation matrix B between roadways and nodes; real fluid data (including pressure matrix P' and flow matrix Q' measured by sensors); roadway length matrix l; pipe diameter matrix d; pipe friction coefficient matrix f; and pipe cross-sectional area matrix s.

[0054] See the simplified diagram of the mine ventilation system. Figure 4 Mine ventilation network diagram , For a set of nodes, ; For the collection of alleyways, ; AE is the lane adjacency matrix, when the lane and alleyway When adjacent, ,otherwise .

[0055] S2: Randomly select the fault branch and set the fault type and fault degree, and randomly generate the fault wind resistance. Based on the wind resistance before and after the fault, perform simulation calculations on the initial pressure, fault pressure, initial flow rate, and fault flow rate before and after the anomaly.

[0056] In this embodiment, simulation calculations are performed on the fluid data prior to the anomaly: In the formula, To simulate the flow rates of all roadways, we have q1, q2, ..., q10 composition, m 3 / s; P is the simulated pressure of all nodes, consisting of p1, p2,..., p8, pa; A E is the roadway adjacency matrix; P' is the perceived pressure matrix; Q' is the perceived flow matrix; d is the roadway pipe diameter matrix, m; f is the roadway friction coefficient matrix; l is the roadway length matrix, m; s is the roadway cross-sectional area matrix, m; and F is the network solving algorithm.

[0057] The fluid data after the anomaly is simulated and calculated, In this embodiment, the mine ventilation network anomaly is set as shown in the following formula: In the formula, B represents the mine ventilation network anomaly setting matrix; E represents each branch, ; and the X vector represents the anomaly type, xi = 0 in the ventilation network indicates no anomaly, xi = 1 indicates that the air door is abnormally opened, xi = 2 indicates that the air door is abnormally closed, and xi = 3 indicates that the roadway is caving; the A vector represents the anomaly degree, , The value range is 0~1; Then, the fluid data under the abnormal condition is simulated and calculated: The mine ventilation network solving calculation model is as follows: wherein, L 1 represents the node mass conservation equation; L 2 represents the node component conservation equation, which can be ignored since the ventilation network does not involve component conservation; L 3 represents the resistance law; r ij represents the branch resistance between node i and node j; L 4 represents the density expression of node i based on the gas state equation. Substitute L 4 into L 1 to obtain the mine ventilation fluid network model L = (L 1 ,L 3 ) The simulation calculation result is obtained by iteratively solving the fluid network model L by the Newton iteration method.

[0058] S3: Use the subnet calculation module to calculate subnets for the mine ventilation network, which are used for knowledge integration and anomaly localization reasoning in the model.

[0059] The random walk algorithm is used to randomly visit branches, and the visited branch is taken as the core branch e. x Connect the core branch with its two endpoints , and the first-order neighborhood of the two endpoints. , The adjacent edges are incorporated into the subnetwork, ultimately forming a star topology T centered on the core branch. S (e x ).

[0060] In this embodiment, the formula for calculating the random access probability of a roadway branch is as follows: In the formula, Indicates a lane The number of adjacent lanes; Represents the sum of the number of adjacent lanes for all lanes; represents a lane visited randomly. The probability of.

[0061] The probability of random branch access in this embodiment is as follows: S4: Use the data management module to label the normal and abnormal data, and store the fluid data and subnet data in normal and abnormal states and their corresponding abnormal labels as training data in the database.

[0062] The pressure and flow data are calibrated based on the simulation results under normal and abnormal conditions in S2, and the fluid data under normal and abnormal conditions obtained in S2, the sub-network obtained in S3, and the fault labels are stored in the database as training data.

[0063] S5: The knowledge integration module reads the training dataset, and through feature integration and autonomous learning, trains the optimal anomaly source localization model M. The optimal anomaly source localization model M is updated into the knowledge model of the anomaly localization module and used for anomaly source localization in the fluid network through centralized deployment.

[0064] In this embodiment, topological feature integration and data feature integration are performed using the following formulas respectively.

[0065] In the formula: A 1 represents the topological feature integration matrix; B The association matrix representing branches and nodes;D represents a node degree matrix; I represents an identity matrix; A E represents a branch adjacency matrix. F 0 represents a fluid network feature; W 0 represents a learnable parameter matrix; diag represents diagonalization; δ represents a nonlinear transformation; F 1 represents integrated features.

[0066] wherein: A 2 represents a data feature integration matrix; W 、 W 1, θ represents a learnable parameter matrix; f represents a column of a feature matrix; F 2 represents integrated features; i 、 j 、 k represents a label.

[0067] S6: Collect mine ventilation network fluid data through a sensor, import the data into an anomaly positioning module in real time via a data management module, output whether an anomaly occurs and an anomaly position, verify and handle the anomaly. Upload and store the data and labels thereof, and update a database.

[0068] In this embodiment, anomaly label prediction is performed by the following formula: wherein, represents a fault prediction label output by a model; M represents an optimal anomaly source positioning model obtained by training; AE represents a roadway adjacency matrix; respectively represent real pressure data perceived at t+1 and t time points; respectively represent real flow data perceived at t+1 and t time points; TS is subnetwork data; d is a roadway pipe diameter matrix; f is a roadway friction coefficient matrix; l is a roadway length matrix; s is a roadway cross-sectional area matrix.

[0069] Then, the obtained anomaly label prediction is verified, and the verified data and labels are transmitted to a data management module to update a training database.

[0070] S7: Continuously repeat S1-S6 to realize real-time positioning of mine ventilation network anomaly sources and dynamic generalization integration of anomaly source positioning knowledge.

[0071] Embodiment 3: Medical gas supply network anomaly positioning A medical gas system is a complete set of facilities that provides necessary gases for patients and medical equipment or processes waste gases. Common medical gas systems include oxygen systems, nitrogen systems, nitrous oxide systems, argon systems, carbon dioxide systems, compressed air systems, negative pressure suction systems, and anesthetic waste gas exhaust systems, etc., and their main structure is a tree network. During the operation of the medical gas system, common abnormal situations include leakage, blockage, insufficient pressure or concentration, etc. Leakage can be caused by pipe aging, poor sealing or loose joints; blockage is usually caused by foreign matter or impurities accumulation in the pipe or filter system failure; and insufficient pressure or concentration can be caused by insufficient gas supply or excessive consumption. These abnormalities will directly affect the supply and use of gases, endangering medical operations and patient safety. In order to effectively locate these abnormalities, the present embodiment will integrate and locate abnormal knowledge through the following process. As Figure 5 is an oxygen transport system diagram, including a pressure pump 5-1, a transport pipeline 5-2, and a valve 5-4.

[0072] S1: Collect basic data of medical gas transport network, including topological relationship and fluid data.

[0073] In the present embodiment, the data obtained includes: branch adjacency matrix A E , branch and node association matrix B, real fluid data (including sensor measured pressure matrix P', flow matrix Q', oxygen concentration matrix c'), pipe length matrix l, pipe diameter matrix d, pipe friction coefficient matrix f, and pipe cross-sectional area matrix s.

[0074] The oxygen transport system diagram is shown in Figure 5 . The oxygen transport network diagram is , is a node set, ; is a branch set, ; AE is a branch adjacency matrix, when branch is adjacent to branch , , otherwise .

[0075] S2: Randomly select a fault branch and set the fault type and fault degree, and randomly generate a fault wind resistance, and simulate and calculate the initial pressure, fault pressure, initial flow, and fault flow before and after the fault according to the wind resistance before and after the fault.

[0076] Simulate and calculate the fluid data before the anomaly: In the formula, Q is the simulated flow of all pipes, composed of q1, q2,..., q 15 m 3 / s; P represents the pressure at all nodes obtained from the simulation, defined by p1, p2, ..., p 16 Composition, pa; c represents the oxygen concentration of all pipes obtained from the simulation, composed of c1, c2, ..., c 15 Composition, %; A E denoted as the pipe adjacency matrix; P' as the sensed pressure matrix; Q' as the sensed flow rate matrix; c' as the sensed oxygen concentration matrix; r as the tunnel diameter matrix (m); f as the friction coefficient matrix for each tunnel; l as the length matrix for each pipe (m); s as the cross-sectional area matrix for each pipe (m); and F as the network solution algorithm.

[0077] Simulation calculations were performed on the fluid data after the anomaly. In this embodiment, the abnormal settings for the medical gas network are represented by the following formula: In the formula, B represents the abnormal configuration matrix of the oxygen transport network; E represents each branch, The X vector represents the anomaly type. In the medical gas network, xi=0 represents no anomaly, xi=1 represents blockage, xi=2 represents leakage, and xi=3 represents insufficient concentration; vector A represents the degree of anomaly. , The value range is 0 to 1; Then, simulation calculations are performed on the fluid data under abnormal conditions: The oxygen transport network solution model is as follows: in, L 1 Represent the equation for the conservation of nodal mass; L 2 Represents the oxygen component conservation equation at the nodes; L 3 This represents the law of resistance; r ij This represents the branch resistance between node i and node j; L 4 This represents the density expression for node i obtained based on the gas law. L 4 Substitution L 1 and L 2 In this process, an oxygen transport fluid network model was obtained. L = (L 1 ,L 2 ,L 3 )The simulation calculation result is obtained by iteratively solving the fluid network model by Newton iteration method. L

[0078] S3: The subnet calculation module is used to calculate different fluid network calculation subnets, which are used for knowledge integration and abnormality positioning reasoning of the model.

[0079] Based on the random walk algorithm, a branch is randomly accessed as a core branch e x , and the core branch is connected with its two end nodes , , and the first-order neighborhood , and adjacent edges of the two end nodes, and finally a star topology T S (e x ) is formed with the core branch e x as the center.

[0080] In this embodiment, the pipe branch access probability calculation formula is as follows: In the formula, denotes the number of adjacent branches of branch e ; denotes the sum of the number of adjacent branches of all branches; and denotes the probability of randomly accessing branch e .

[0081] The random branch access probability of this embodiment is as follows: S4: The data management module is used to label the normal and abnormal labels of data, and the fluid data, subnet data in normal and abnormal states and their corresponding abnormal labels are stored into the database as training data.

[0082] According to the simulation calculation results in the normal and abnormal states in S2, the pressure, flow rate and concentration data are labeled, and the fluid data in the normal and abnormal states obtained in S2, the subnet obtained in S3 and the fault label are stored into the database as training data.

[0083] S5: The knowledge integration module reads the training data set, and trains the optimal abnormal source positioning model M through feature integration and autonomous learning. The optimal abnormal source positioning model M is updated to the knowledge model of the abnormal positioning module, and is used for fluid network abnormal source positioning through centralized deployment.

[0084] In this embodiment, the topological feature integration and data feature integration are respectively performed by the following formulas ​​ wherein: A 1 represents the topological feature integration matrix; B 0 represents the branch and node association matrix; D 1 represents the node degree matrix; I 0 represents the unit matrix; A E 0 represents the branch adjacency matrix. F 0 represents the fluid network feature; W 0 represents the learnable parameter matrix; diag represents diagonalization; δ 0 represents the nonlinear transformation; F 1 represents the integrated feature.

[0085] wherein: A 2 represents the data feature integration matrix; W , W 1, θ 0 represents the learnable parameter matrix; f 0 represents a column of the feature matrix; F 2 represents the integrated feature; i , j , k 0 represents the label.

[0086] S6: The medical gas network fluid data collected by the sensor is imported into the anomaly positioning module in real time via the data management module, and whether an anomaly occurs and the anomaly position are output, the anomaly is verified and processed. The data and labels are uploaded and stored, and the database is updated.

[0087] In this embodiment, the anomaly label prediction is performed by the following formula: wherein, 0 represents the fault prediction label output by the model; M represents the optimal anomaly source positioning model trained; AE 0 represents the branch adjacency matrix; respectively represent the real pressure data perceived at t+1 and t time; respectively represent the real flow data perceived at t+1 and t time; respectively represent the oxygen concentration data perceived at t+1 and t time; TS is the sub-network data; d is the roadway pipe diameter matrix; f is the friction coefficient matrix; l is the pipe length matrix; s is the branch cross-sectional area matrix.

[0088] Then, the obtained anomaly label prediction is verified, and the verified data and labels are transmitted to the data management module to update the training database.

[0089] S7: constantly repeating S1-S6, realizing real-time positioning of the abnormal source of the medical gas transportation network and dynamic generalization integration of the abnormal source positioning knowledge.

[0090] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A fluid network anomaly source localization and knowledge integration method, characterized in that, Comprise: S1: Collect fluid network basic data, including topological relationship, branch parameter and fluid data; Topological relations: Branch-adjacency matrix A E , Node-branch incidence matrix B; Branch parameter: branch length l, diameter d, friction coefficient f, cross-sectional area s; Fluid data: pressure, flow, concentration data measured by sensors; S2: randomly select a fault branch on the fluid network and set the fault type and fault degree, and randomly generate the fault wind resistance According to the wind resistance before and after the fault, the pressure data, flow data, and concentration data before and after the anomaly are simulated and calculated. S3: Calculate the fluid network to divide into several sub-networks for knowledge integration and abnormal positioning reasoning of the model; S4: Calibrate the normal and abnormal labels of pressure data, flow data and concentration data, and store the fluid data, sub-network data and label data into the database as training data set; S5: Train the training data set to obtain the optimal abnormal source positioning model for fluid network abnormal source positioning; S6: Collect fluid network fluid data through sensors, output whether an abnormality occurs and the abnormal position, verify and handle the abnormality; S7: Continuously repeat S1-S6 to realize real-time positioning of fluid network abnormal source and dynamic generalization integration of abnormal source positioning knowledge.

2. The fluid network anomaly source localization and knowledge integration method of claim 1, wherein, In the step S2, the mathematical model of simulation calculation is: where Q is the simulated total lane flow, consisting of q1, q2,..., qn; P is the simulated total node pressure, consisting of p1, p2,..., pn; c is the simulated total lane gas concentration, consisting of c1, c2,..., cn; A is the lane adjacency matrix; m is the lane length matrix, m; f is the lane friction factor matrix; and l is the lane length matrix, m. n where Q is the simulated total lane flow, consisting of q1, q2,..., qn; P is the simulated total node pressure, consisting of p1, p2,..., pn; c is the simulated total lane gas concentration, consisting of c1, c2,..., cn; A is the lane adjacency matrix; m is the lane length matrix, m; f is the lane friction factor matrix; and l is the lane length matrix, m. 3 where Q is the simulated total lane flow, consisting of q1, q2,..., qn; P is the simulated total node pressure, consisting of p1, p2,..., pn; c is the simulated total lane gas concentration, consisting of c1, c2,..., cn; A is the lane adjacency matrix; m is the lane length matrix, m; f is the lane friction factor matrix; and l is the lane length matrix, m. m where Q is the simulated total lane flow, consisting of q1, q2,..., qn; P is the simulated total node pressure, consisting of p1, p2,..., pn; c is the simulated total lane gas concentration, consisting of c1, c2,..., cn; A is the lane adjacency matrix; m is the lane length matrix, m; f is the lane friction factor matrix; and l is the lane length matrix, m. E where Q is the simulated total lane flow, consisting of q1, q2,..., qn; P is the simulated total node pressure, consisting of p1, p2 The calculation model of network solution algorithm F is as follows: wherein, L 1 represents a node mass conservation equation; L 2 represents a node component conservation equation, used in a mixed gas flow network; L 3 represents a resistance law; r ij represents a branch wind resistance between node i and node j; L 4 represents a density expression of node i based on a gas state equation; substituting L 4 into L1 and L 2 , a fluid network model L=(L 1 ,L 2 ,L 3 ) is obtained; and a simulation calculation result is obtained by iteratively solving the fluid network model L by a Newton iteration method.

3. The fluid network anomaly source localization and knowledge integration method of claim 1, wherein, The step S3 comprises: S31: For the fluid network , is a set of nodes, ; is a set of lanes, ; AE is a lane adjacency matrix, when a lane is adjacent to a lane , , otherwise ; The random lane access probability is obtained by the following formula: wherein represents the number of adjacent roadways of a roadway ; and represents the sum of the number of adjacent roadways of all roadways; represents the probability of a random access roadway ; S32: Randomly select the core tunnel based on the random walk algorithm. Connect the core tunnel with its two end nodes , Incorporate them into the subnetwork; at the same time, include the first-order neighborhoods of the nodes at both ends. , The adjacent lanes are also included in the sub-network; ultimately forming a network centered on the core lanes. The core is a star-shaped topology TS(ex), and the core lanes and their associated local networks together constitute the main structure of the subnetwork.

4. The fluid network anomaly source localization and knowledge integration method of claim 1, wherein, The step S4 comprises: S41: According to the simulation calculation results under normal and abnormal conditions in step S2, calibrate the fault labels of pressure, flow and concentration data; S42: Store the fluid data under normal and abnormal conditions obtained in step S2, the sub-network obtained in step S3 and the corresponding labels into the database as training data.

5. The fluid network anomaly source localization and knowledge integration method of claim 1, wherein, The step S5 comprises: S51: Topological feature integration, construct a first-order neighborhood prior feature integration with weight based on the topological relationship of the fluid network; S52: Data feature integration, construct a first-order neighborhood posterior feature integration with weight based on the fluid network data; S53: Perform autonomous learning training of the abnormal positioning model.

6. The fluid network anomaly source localization and knowledge integration method of claim 5, wherein, The step S52 comprises: Construct a cross-entropy loss function for fault diagnosis: In the formula: is the cross-entropy loss function for fault diagnosis; C g is the number of categories; is the true label of the fault; is the predicted label of the fault; is the topological feature, i.e., the lane adjacency matrix of the ventilation network; X is the data feature; Take the data features, topological features and fault labels of the sub-network as the input of the model; After topological feature integration and data feature integration, connect the features, input the Softmax layer to calculate the predicted label, and the label dimension is lane number x fault type number; Calculate the cross-entropy loss of the predicted label and the real label, and realize autonomous learning by back propagation, so as to realize autonomous learning training of the abnormal positioning model.

7. A fluid network anomaly source localization and knowledge integration system, comprising: Comprise: Sub-network calculation module for completing the step S3 of dividing the fluid network into several sub-networks; Simulation module for completing the simulation in step S2; Knowledge integration module for completing the autonomous learning training of the abnormal source positioning model in step S5; Abnormal positioning module for completing the positioning of the fluid network abnormal source in step S6; Data management module for importing, storing, exporting, generating and calibrating the topological relationship in step S1, the fluid data in step S2 and the label data in step S4.

8. The fluid network anomaly source localization and knowledge integration system of claim 7, wherein, The sub-network calculation module comprises a random sub-network dynamic calculation module and a fixed sub-network module; The random sub-network dynamic calculation module obtains several sub-networks; The fixed sub-network module obtains fixed sub-networks by manual input.

9. The fluid network anomaly source localization and knowledge integration system of claim 7, wherein, The knowledge integration module comprises: A first data import module, which receives stored topological relations, fluid data and label data in the data management module; A feature integration module, which is used to complete topological feature integration and data feature integration An autonomous learning module, which is used to train the abnormal source positioning model.

10. The fluid network anomaly source localization and knowledge integration system of claim 7, wherein, The abnormal positioning module comprises: A second data import module, which imports fluid data, topological relations and sub-networks from the data management module as inputs of the abnormal positioning model; A knowledge model module, which is the optimal abnormal source positioning model trained by the knowledge integration module; An abnormal output module, which is responsible for finally transcoding and outputting the position of the abnormality according to the calculation result of the optimal abnormal source positioning model.

Images